RE: VM start time slows with large number of VMs

[Mats Petersson <mats@planetcatfish.com>]

At 17:27 11/09/2007, Carb, Brian A wrote:

>  I tried timing xenstore-ls and then tried it again with the shared 
> memory idea (adding the 2 lines to the /etc/init.d/rc3.0/S10xend 
> script before it starts xend). Here are the average times for xenstore-ls:
>
>  # vms     std    shm
>001-020    .67    .59
>021-040   1.70   1.54
>041-060   2.41   2.36
>061-080   3.26   3.14
>081-100   4.06   3.95
>101-120   4.96   4.90
>121-140   6.56   6,02
>141-160   7.33   7.21
>161-180   8.00   8.62
>181-200   9.90   9.61
>
>Without using shared memory, the total elapsed time to start 200 VMs 
>is about 95 minutes, vs 81 minutes with the shm statements. Time to 
>start the 200th vm, from create to unpaused, is about 55 seconds, 
>and with xenstore in shared memory that time is about 43 seconds. 
>I'm not seeing the 200+ seconds I was seeing before, but that time 
>was for vm263 and I've only been able to test with 200 VMs. This 
>week I'll try to test with more.
>
>Since the shared and non-shared memory results are so close, how can 
>I verify that the xenstore data is actually going in and out of 
>shared memory? It seems as if the start times became reduced once I 
>cleared out the old info from xenstore, regardless of shared memory.


That would make sense.

I'm not sure how well the xenstore structure is designed, with 
respect to holding hundred and thousands of different entries - I had 
a quick look at it, but not really looking at it deeply, just very 
briefly - the code I was looking at seemed to do a lot of small 
function calls to dig data out of the database a byte or some few 
bytes at a time, which didn't make me say "Wow, that's clever".

--
Mats


>Is it worth trying bootloader/bootentry instead of kernel/ramdisk as well?
>
>brian carb
>unisys corporation - malvern, pa
>
>-----Original Message-----
>From: Daniel P. Berrange
>Sent: Wednesday, September 05, 2007 1:38 PM
>Subject: Re: [Xen-devel] VM start time slows with large number of VMs
>
>On Wed, Sep 05, 2007 at 06:11:06PM +0100, Keir Fraser wrote:
> > xenstore slowness seems a reasonable hypothesis. How does performance
> > of Œtime xenstore-ls >/dev/null¹ change as you increase the 
> number of guests?
>
>There's another issue with the way hotplug scripts do locking which 
>exhbits fairly pathelogical behaviour when a large number of guests 
>are started concurrently. We've got a patch we're evaluating which 
>changes  the use of lock files + looping + sleeps, to instead use 
>the 'flock' shell util.
>
>I suspect the performance issues will be a combination of both the 
>locking stuff and the xenstore I/O. You might also want to try 
>putting /var/lib/xenstored onto tmpfs eg in  initscripts before 
>xenstored starts ...
>
>   mkdir /dev/shm/xenstored
>   mount --bind /dev/shm/xenstored /var/lib/xenstored
>
>
>Dan.
>
>_______________________________________________
>Xen-devel mailing list
>Xen-devel@lists.xensource.com
>http://lists.xensource.com/xen-devel